The AI Module Profiles standard

The AI Module Profiles (MPAI-PRF) standard specifies a standard way to signal an AI Module Profile: a label that uniquely identifies the set of Attributes (input data, output data and functionality) that characterises an AI Module (AIM) instance.

AIMs with the same name may receive more or less input, and produce more or less output, depending on the specific features of the AIM. For example, the Natural Language Understanding AIM (MMC-NLU) of Conversation with Emotion receives Recognised Text and produces Refined Text and Meaning. The MMC-NLU of Communicating Entities in Context may also use Audio-Visual Scene Descriptors and Object Instances as input to understand a sentence better. Since it is not realistic to require every MMC-NLU to use all of this, MPAI-PRF lets an implementation state exactly what the AIM supports, in a form that is both machine-readable and understandable by a human.

The Structure of a Profile

An AIM Profile is identified by:

  • Three characters identifying the Technical Specification where the AIM is specified.
  • Three characters identifying the AIM in that Technical Specification.
  • The Version and Subversion of the Technical Specification.
  • The sequence of coded Attributes of the input and out data that characterises the Profile.

Each Attribute has a three-character code, e.g. TXO for Text Object, AVM for Avatar Model, EPS for Personal Status and TRN for Translation.

Main Functions of AI Module Profiles

The MPAI-PRF standard enables the following essential functions:

  • Identification – Names the specification, AIM and version an implementation conforms to.
  • Capability Signalling – States which Attributes the implementation supports.
  • Refinement – States which Sub-Attributes of an Attribute are supported, such as the modalities of Personal Status or the language pairs of a translation.
  • Compact Signalling – Allows a capability set to be listed by what is supported or by what is not, whichever is shorter.
  • Interface Declaration – Tells users which interfaces the implementation exposes.

These functions allow a Module to select, and an integrator to verify, an AIM that has exactly the capabilities that the application needs.

The AIMs Covered

MPAI-PRF currently specifies Profiles for eight AIMs:

  • Audio-Visual Scene Rendering (PAF-AVR).
  • Entity Context Understanding (HMC-ECU).
  • Entity Dialogue Processing (MMC-EDP).
  • Natural Language Understanding (MMC-NLU).
  • Personal Status Display (PAF-PSD).
  • Personal Status Extraction (MMC-PSE).
  • Text and Speech Translation (MMC-TST).
  • Text-to-Speech (MMC-TTS).

The list will be revised when existing AIMs come to require Profiles, when their functionality is extended, or when new AIMs requiring Profiles are specified. Basic and Composite AIMs are treated uniformly.

Operation Model

A Profile can be signalled in two ways:

  • ALL followed by the Attributes, each prefixed by “–” in the parentheses, that are not supported. For example, MMC-NLU-V2.1(ALL-AVS-OII) is an NLU that supports everything except scene and object identifrication.
  • NUL followed by the Attributes, each prefixed by “+” in the parentheses, that are supported. For example, MMC-NLU-V2.1(NUL+TXO+TXR) is an NLU that supports only Text Object and Recognised Text.

Sub-Attributes refine an Attribute. They follow the Attribute prefixed by “@”, each prefixed by “#” (underlined in the following two examples):

  • PAF-PSD-V1.1(ALL@EPS#PSS#PSF) supports all Attributes, with Personal Status limited to Speech and Face.
  • MMC-TST-V2.1(NUL+TXO@TRN#eng->ita) supports only text translation from English to Italian.

The capabilities of an AIM thus have two dimensions: its Profile, given by its Attributes, and its Level, given by its Sub-Attributes. An AIM need not expose the interfaces of Attributes it does not support. If it supports all Attributes, it shall expose all interfaces.

Powered by the MPAI AI Framework

AIMs identified by MPAI-PRF Profiles are executed in the MPAI AI Framework (MPAI-AIF), which provides:

  • A modular and interoperable execution environment for AIMs.
  • Configuration and orchestration of AIMs in a Module.
  • Platform-independent implementation.

A Profile lets the Framework, and whoever composes a Module, know which implementation of an AIM fits which Module.

Open and Flexible Design

MPAI-PRF follows key design principles:

  • Implementations are not forced to support every capability of an AIM.
  • Capabilities are declared unambiguously, in a label readable by machines and people.
  • Two dimensions, Profile and Level, cover both what an AIM handles and how far.
  • The list of AIMs and Attributes grows as MPAI specifications evolve.

Benefits

The MPAI-PRF standard will support:

  • Implementers to offer AIMs with a subset of capabilities, and say so.
  • Integrators to choose the AIM implementation that matches their application.
  • Service providers to describe precisely what their AI components can do.
  • Conformance testing to check an implementation against the capabilities it declares.
  • Users to know what to expect from an AI component before using it.

A New Paradigm for AI Components

MPAI-PRF promotes a shift from all-or-nothing AI components to:

  • Components that declare exactly what they support.
  • Modules assembled from components matched on declared capabilities.
  • A market of implementations differing in scope as well as in performance.
  • A common, compact vocabulary for AI capabilities.

Conclusion

MPAI-PRF provides an interoperable way to signal the capabilities of AI Modules that:

  • Identifies an AIM, its specification and its version.
  • States the Attributes and Sub-Attributes it supports.
  • Allows compact signalling by inclusion or exclusion.

The standard enables AI Modules of different scope to coexist, be selected and be combined reliably.